A Wild Week in AI: Enzyme Design, 700 Math Breakthroughs, and Room-Temperature Magnets

This week in AI was genuinely absurd, and sorting my notes felt like defusing a firework stand. DeepMind unveiled a from-scratch enzyme designer, OpenAI dropped hundreds of mathematics manuscripts, and researchers surfaced room-temperature magnetic semiconductor candidates. A Traictory analysis cautions that the enzyme claims rest on a preprint without peer review. Against this scale, the real bottleneck looks like human capacity to process it all.

The first stop is WorldPlay 2 , which turns a single image into a playable world inside the browser. Walking with WASD, looking with arrow keys, jumping with space, and steering scenes through mid-run text events are all supported. Weights and code landed on GitHub on October 6, 2026, with a browser demo on Reactor following on October 7. An ArtrealmAI report notes the system continues Tencent Hunyuan HY-World 1.5 and holds long-horizon consistency through compressed memory. The turbo variant weighs 69 GB.

Motion Space Flow generates human motion from text prompts and stands out through joint constraints. Alongside moves like a high left-leg kick or crawling, you can pin a hand or foot position and let the animation grow around that anchor. The release includes training and evaluation scripts besides the code, totaling only about 1.1 GB. The Apache 2.0 license opens the door to commercial experiments. At a size that fits consumer GPUs, independent animators get a serious new tool.

On the open video front, Kandinsky 6.0 ships in two sizes: a 3.8-billion-parameter Lite and a 29-billion-parameter Pro. Both render 5-second clips with synchronized 44 kHz audio including lip-sync, supporting text-to-video and image-to-video but not reference-to-video. A companion latent upscaler multiplies output resolution up to four times. The KandinskyLab page confirms Diffusers and ComfyUI support. Lite is 6.38 GB while Pro reaches 60 GB.

Generative interfaces and language models

The OpenAI item that caught my eye is intelligent UI generation. ChatGPT can now answer with clickable interactive panels instead of walls of text, for example a diagram of a bicycle whose parts you select to learn how it works. The shift turns the chatbot from document reader into application. More visual and interactive answers move technical explanation from memorization toward hands-on use. An OpenAI notice says the feature is rolling out gradually.

Amazon answered with ALoDLM , an adaptively looped diffusion language model marrying diffusion with LLM design. Instead of emitting tokens strictly left to right, the model predicts many answer fragments in parallel and refines them over recurrent loops, each loop acting like extra thinking time. The HuggingFace paper summary reports top scores on reasoning, mathematics, science, and coding against similar-size models. Parallel generation also cuts latency. This Frankenstein design could reshape inference economics.

The biotech bomb came from DeepMind as AlphaProtein Novo . Rather than mutating natural enzymes, the system designs the 3D structure and amino-acid sequence for a target reaction from scratch, then checks folding with AlphaFold 3. Lab tests produced a transferase assembling piperidine, a ring common in approved drugs, plus an esterase degrading DEHP, a plasticizer pollutant. A Traictory review stresses the work is an October 5 bioRxiv preprint without peer review. Early days, but the direction is set.

A historic day for mathematics

Plant genetics starred Gemma 4 paired with Botanic-1 . Gemma 4 plays scientist, drafting hypotheses, writing pipelines, and running analysis, while Botanic-1 brings expertise from training on DNA sequences of 320 plant species. Given roughly 2,500 candidate mutations, the system ranked the true melon-flower variant at the top. A Google Blog note says the embedding models share similar infrastructure. Candidate screening for field trials could shrink dramatically.

The OpenAI mathematics release was the headline: 722 manuscripts from an unreleased frontier model, around 4,000 attempted problems, and roughly 3 hours of compute per problem. Results sit in a GitHub repository, with release practices shaped with the independent advisory group on mathematics and AI at the IAS. An OpenAI page says revision and citation protocols will run with the community. Hundreds of proofs in a single day starts a new era that strains reviewer capacity.

Four highlights tower above the rest. A quasi-Riemann hypothesis advance opens a new front on the zeta function and prime distribution. On Hilbert tenth over rationals , a proof claims the decision method proven impossible for integers in 1970 stays impossible even allowing fractions. That closes a door open for decades. Independent verification is pending, so celebration is premature, yet the scope is breathtaking.

The same bundle claims the matrix-multiplication exponent falls from 2.37 to 2.25 . The gap looks small, yet savings compound across billions of repeated multiplications, cheapening training and inference directly. Partial progress on the Birch and Swinnerton-Dyer Millennium conjecture covers rational points on elliptic curves. Should both survive review, textbooks get rewritten. Together with the DeepMind enzyme work, the week spanned the full arc from theory to practice.

Open models and search infrastructure

Google released Embedding Gemma 2 , an open embedding model mapping text, images, audio, and video into one unified space. Text-only runs need just 270 million parameters, with 170 million more for vision and 300 million for audio, keeping total size near 1.5 GB. A Google Blog announcement says the first generation passed 20 million downloads and powers on-device search plus privacy-first retrieval pipelines. The Apache 2.0 license eases enterprise adaptation. A small but mighty infrastructure piece.

Mistral, the well-known European lab, shipped Large 4 , unofficially Le Chonk. The trillion-parameter mixture-of-experts activates only 52 billion parameters per use and supports native multimodality. A Mistral announcement says the preview API is live on Mistral Studio with weights dropping at month end. Yet Artificial Analysis ranks it behind Kimi 3, GLM 5.3, and DeepSeek Flash variants, and pricing draws criticism. Judgment should wait for the open weights.

Meta contributed GenIA , turning a photo or video into complete 3D objects. On top of a SAM 3D base, a test-time alignment loop keeps checking the generated object against observed geometry and corrects appearance and pose from the differences. The arXiv paper explains missing viewpoints get completed through generative priors. Code is public under a non-commercial Creative Commons license. A practical path for product visualization.

Two room-temperature magnet candidates from the Vals team were the hardware surprise. A crew of Claude Opus 5.5 agents found two antiferromagnetic semiconductor candidates with zero net magnetism that still sort electrons by spin: one newly designed compound, one material first synthesized in 1999. A Vals report says all calculations and code are shared alongside known caveats. Denser packing and faster switching could reshape memory architectures. Experimental confirmation is still required.

Robotics, 3D, and image models

FlashDex Retarget converts human hand demonstrations into robot-hand motion with a single policy trained across many movements, peeking at upcoming frames to anticipate. Benchmarks beat prior methods on average, and code is open for XHand and Sharp Wave hands. Training needs around 110 GB of GPU memory, which strains home users. Still, it shortens the bridge from demonstration to deployment with clean engineering. A similar simplification is happening in images through the Speridlabs pixel-space bet.

The 3-billion-parameter Iris 3B generates directly in pixel space with no VAE, sidestepping lossy compression and texture-biased latents. Text-to-image and image-to-image are supported, though semantic editing in the Nano Banana style is absent. A Speridlabs gallery shows photorealistic frames plus confident large-type rendering. The 12 GB base model accepts fine-tuning, with a depth-estimation example shared. Behind the leading open models, but a light and honest alternative.

Terra teaches simulated bodies muscle-driven locomotion : over 9 hours of motion data and 18 demonstrations yield correct steps on slopes and stairs, extended by specialist policies for jumps and rolls. Code is open under Apache 2.0. Codim Recon converts ordinary photos into editable physics-ready 3D scenes, modeling cables as flexible curves, bags as surfaces, and cushions as deformable solids, then testing them through robot interaction. Its code is coming soon. Together they close the simulation gap.

Two ready-to-use releases close the week. Google quietly shipped Nano Banana 2.1 , promising more accurate generation and editing for complex instructions, multi-character scenes, and in-image text, with up to 14 reference images and 4K support. A DeepMind product page shows the family splitting into Lite, Flash, and Pro options. Meanwhile Anthropic landed Claude Haiku 5.5 , beating its predecessor by a wide margin on knowledge work and agentic coding at roughly 75 percent lower operating cost. The bottleneck now is our reading capacity, not compute.